Lance Format
Lance is an open-source lakehouse format for multimodal AI. It combines a file format, table format, and catalog specification to provide fast random access and scans for vector and full-text search, feature engineering, and model training on object storage, while retaining SQL analytics, ACID transactions, time travel, and open-engine integrations.
At a Glance
- AI/ML engineering teams
- Data-platform and lakehouse teams
- Teams building vector databases and RAG systems
- LLM, vision, robotics, and autonomous-vehicle researchers
- +2 more
AI Tools by Lance Format
(1)
Lance
Open Source Multimodal AI Lakehouse
Discussions
No discussions yet
Be the first to start a discussion about Lance Format
Latest News
Lance datasets used directly from object storage for LeRobot training; hybrid data mining and distributed search reached 1,066 QPS across 10B vectors.
LanceDB newsletter highlighted Lance-related GPU data-loading work, CrewAI cognitive memory on LanceDB, and the launch of the Reverie AI-builder summit.
Multi-bit RaBitQ reached 96% recall, Volcano Engine reduced training pipelines from seven days to one, and Lance powered 100 Hz robotics data ingestion.
Lance was benchmarked against Delta Lake and Iceberg for S3 metadata performance; blob-row updates and direct world-model training from object storage were highlighted.
Products & Services
An Arrow-based, AI-oriented columnar storage format optimized for fast random access, scans, vector data, multimodal blobs, embeddings, and indexes.
Table-layer capabilities including ACID transactions, snapshot/versioning and time travel, schema evolution, updates/deletes, and secondary indexes for vector, full-text, and SQL queries.
A catalog/namespace specification with Rust, Python, and Java SDK code generation plus implementations and integrations for catalogs such as Apache Hive, Apache Polaris, Apache Gravitino, Unity Catalog, and AWS Glue.
Open-source Rust core SDK with Python and Java bindings for creating, reading, updating, indexing, versioning, and querying Lance datasets.
Market Position
Lance positions itself as an AI-native alternative/complement to general-purpose formats such as Apache Parquet, Apache Iceberg, and Delta Lake: unlike formats primarily optimized for analytical scans or table management, it makes random access, vectors, embeddings, indexes, and multimodal blobs first-class while still supporting SQL, ACID transactions, and lakehouse interoperability. The project is open source under Apache-2.0 and is designed to work with existing engines and catalogs rather than requiring a closed platform.
Leadership
Founders
Chang She
Co-creator of Pandas; also identified in public materials as co-founder and CEO of LanceDB, the organization stewarding Lance.
Lei Xu
Contributor to Apache HDFS; identified by industry coverage as a creator of the Lance format.
Founding Story
Lance was introduced in 2022 to make high-dimensional and multimodal data easier and faster to query than traditional columnar formats, particularly for vector retrieval and workloads combining embeddings with structured data, images, and other large binary objects.
Target Markets
- AI/ML engineering teams
- Data-platform and lakehouse teams
- Teams building vector databases and RAG systems
- LLM, vision, robotics, and autonomous-vehicle researchers
- Organizations operating multimodal datasets on cloud object storage
- Open-source data infrastructure and query-engine communities
- Multimodal AI lakehouses
- Vector search and retrieval-augmented generation (RAG)
- LLM and vision-model training
- Feature engineering and materialized model features
- Large-scale embedding and image/video dataset management
- Agent memory and multimodal context lifecycle management